Reinforcement learning for real-time luminosity optimization in colliders

This work introduces a reinforcement learning algorithm designed for real-time luminosity optimization in collider experiments. The neural network architecture is selected from multiple candidates through systematic evaluation of their training performance. Prior to training, the input data undergo multistage preprocessing before being fed into the neural network. Our approach combines off-line pretraining on historical accelerator data with online fine-tuning during operation. By processing accelerator measurements over multisecond timescales, the reinforcement learning model dynamically adjusts the magnetic structure to maintain luminosity stability under changing beam conditions. The autonomous nature of the method eliminates the need for manual intervention, enhancing both operational efficiency and beam stability in long-term operation. Experimental validation on the VEPP-4M collider demonstrates the feasibility of the approach and provides a foundation for future development and deployment in accelerator systems.